Finance and Economics Discussion SeriesDivisions of Research & Statistics and Monetary Affairs

Federal Reserve Board, Washington, D.C.

Post-crisis Signals in Securitization: Evidence from Auto ABS

Elizabeth Klee and Chaehee Shin

2020-042

Please cite this paper as:Klee, Elizabeth, and Chaehee Shin (2020). “Post-crisis Signals in Securitization: Evidencefrom Auto ABS,” Finance and Economics Discussion Series 2020-042. Washington: Boardof Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2020.042.

NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminarymaterials circulated to stimulate discussion and critical comment. The analysis and conclusions set forthare those of the authors and do not indicate concurrence by other members of the research staff or theBoard of Governors. References in publications to the Finance and Economics Discussion Series (other thanacknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Post-crisis Signals in Securitization: Evidence from Auto ABS

Elizabeth Klee∗ Chaehee Shin†

March 30, 2020

Abstract

We find significant evidence of asymmetric information and signaling in post-crisis offeringsin the auto asset-backed securities (ABS) market. Using granular regulatory reporting data, weare able to directly measure private information and quantify its effect on signaling and pricing.We show that lenders “self-finance” unobservably higher-quality loans by holding these loansfor longer periods to signal private information. This signal is priced in initial offerings of autoABS and accurately predicts ex-post loan performance. We also demonstrate that our resultsare robust to exogenous shifts in the demand and supply of auto loans. Despite an environmentof post-crisis enhanced transparency and securitization standards, signaling may be motivatedby inattentive investors and regulations enforcing “no adverse selection” in constructing ABS.

JEL Codes: G23, G14, D82Keywords: Securitization; asymmetric information; signaling; financial regulation; auto loans;securities markets

Acknowledgments: The views expressed here are strictly those of the authors. They do not necessarily representthe position of the Federal Reserve Board or the Federal Reserve System. We thank Craig Furfine, Li Geng, AliHortacsu, Karen Pence, Chiara Scotti, and seminar participants at the Federal Reserve Board for helpful comments.We thank Kevin Kiernan, Jason Lopez, David Smythe, and Steven Urry for research and technical assistance.∗Federal Reserve Board, E-mail: elizabeth.c.klee@frb.gov†Federal Reserve Board, E-mail: chaehee.shin@frb.gov

1

1 Introduction

Shoddy securitization practices were at the heart of the Great Financial Crisis (GFC) (Gorton

and Metrick, 2013). Pre-crisis, securitization was subject to widespread asymmetric information,

with investors having little information on the quality of the underlying loans. This asymmetric

information resulted in adverse selection and an associated “lemons problem” in securitization

markets (Akerlof (1970), Downing, Jaffee, and Wallace (2008)). In some cases, to mitigate the

adverse selection, issuers signaled the quality of the loans (Spence (1973), Adelino, Gerardi, and

Hartman-Glaser (2019)).

In the wake of the GFC, the Congress and regulators demanded new practices aimed at solving

the asymmetric information problem in securitization. In this paper, we explore the effect of two

regulatory changes on asymmetric information and signaling in the auto ABS market. One change

was that the SEC’s Regulation AB was modified to expand the information available to investors,

by requiring the disclosure of detailed information on each loan comprising an auto ABS. Before the

regulatory change, auto ABS issuers were only required to provide aggregate summary statistics

in the prospectus associated with the securities offering.1 The other change was that the major

rating agencies had to attest that the loans comprising the ABS satisfied selected representations

and warranties (“reps and warranties”) common to the particular type of ABS. For example, for

auto ABS, a common rep and warranty was that “no adverse selection” was used in constructing

the pool of auto loans eventually securitized into auto ABS.2 Practically, this new requirement

prohibited issuers from only including “lemons” in securitizations.

Our main result is that, even in a post-crisis regulatory environment, auto ABS issuers signal

private information regarding borrower or collateral quality to investors. In theory, the regulatory

changes should eliminate the need to signal in this market. In practice, we show that this is not the

case. More generally, (1) we directly measure private information; (2) we demonstrate signaling

in a market where adverse selection is expressly forbidden and in an environment of enhanced

transparency; and (3) we observe signaling in post-crisis issuance of auto ABS, a security far

1Refer to Securities and Exchange Commission, 17 CFR Parts 229, 230, 232, 239, 240, 243, and 249, Asset-BackedSecurities Disclosure and Registration, available at https://www.govinfo.gov/content/pkg/FR-2014-09-24/pdf/

2014-21375.pdf.2Refer to Securities and Exchange Commission, 17 CFR Parts 232, 240, 249, and 249b, Nationally Recognized Sta-

tistical Rating Organizations, available at https://www.sec.gov/rules/final/2014/34-72936.pdf. See Appendixfor more details.

2

simpler and and an environment less prone to private information than private-label MBS issued

in the 2000s. Taken together, these results suggest that asymmetric information is inherent in

securitization, and thus the benefits of risk pooling should be weighed against the costs of imperfect

information.

Our direct measure of private information comes from the granular ABS loan-level filings. These

filings include not only “hard” information regarding the borrower and the underlying collateral,

but also, indicators of potential “soft” information that the issuer may have, but is not directly

available to the investor.3 Examples of hard information include variables commonly used in auto

loan underwriting, such as borrower credit scores and payment-to-income ratios, and loan standards

and terms. Examples of soft information include whether there was an underwriting exception —

suggesting that the underwriters may have obtained additional information about the loan, and

whether income was verified — suggesting additional information attained by the lender since

income reporting is usually taken at the borrower’s word.

Importantly, acquiring the indicators of soft information may be relatively costly to the investor.

Specifically, in addition to the granular filings data, all ABS deals provide prospectuses to investors

prior to sale, in which they include pages (“tear sheets”) that disclose summary statistics on the

collateral pool. There are three important distinctions between the tear sheets and the loan-level

data. First, the tear sheet contains some, but not all, variables disclosed on the loan-level filings.

Most often, the prospectuses include summary statistics of hard information, but omit statistics of

soft information. Second, the tear sheets are easy to use, while the granular loan level data require

sophisticated programming skills to extract and analyze, a costly exercise. And third, pre-crisis,

only the tear sheets were available to investors as a source of information; the granular data were

not reported (Neilson et al., 2019). As such, it may be costly for investors to change their process

for evaluating ABS.

The existence of hard and soft information and costly information acquisition for investors

present a conflict for issuers subject to a no-adverse-selection rule. In particular, prices in initial

offerings of securities may not appropriately reflect the quality of the underlying pool. We demon-

strate that issuers surmount this problem by signaling. Consistent with the predictions of Leland

3 Rajan, Seru, and Vig (2015) and Keys, Seru, and Vig (2012) use a similar soft-hard information distinction inthe subprime MBS market.

3

and Pyle (1977), who suggest that entrepreneurs can signal unobservably higher quality projects

to potential investors through self-financing, we show that issuers hold unobservably higher-quality

loans for longer periods. Of note, the longer the issuer holds the loan, the higher the costs of

funding it — a form of self-financing. We find that issuers hold loans with soft information longer

by adjusting the length of time between originating the loan and eventual securitization, a practice

known as “warehousing.” Importantly, the prospectuses generally report the average warehousing

time of the pool, further bolstering our theory that it serves as a signal of quality.

Our results suggest that signaling based on the private, soft information is economically signifi-

cant, while that for the reported, hard information is not. Consistent with the hypotheses, our soft

information indicators significantly predict the signaling variable: loans made under an underwrit-

ing exception experience warehousing times that are 7 to 15 days longer. Loans for which obligor

income has been verified are also warehoused for longer periods. Loans made on used cars shorten

warehousing times by roughly 7 to 8 days. All effects are statistically significant and economically

meaningful. In contrast, warehousing times barely budge for hard information. For instance, a

one standard deviation jump in obligor credit score (96 points) corresponds to only a 0.01 day

increase in the number of warehousing days. Thus, while the effect is statistically significant, it is

far less economically meaningful. We interpret this as evidence that signaling is used primarily to

communicate soft information, not hard information. We conduct various robustness analyses to

address other possibilities; our overall results are unchanged.

Beyond the massive filings information, the structure of the auto market and of auto ABS allows

for other direct tests of signaling based on asymmetric information. We use three “experiments”

that provide plausibly exogenous variation in collateral and borrower quality to gauge their effects

on loan warehousing times. Our tests explore the effects of auto recalls, model updates, and “super

sales” on signaling. Our results conform to our hypothesis regarding signaling: worse collateral

and weaker borrowers suggest shorter warehousing times. We also find that the signal is sensitive

to the cost of producing it. In particular, we illustrate that warehousing times significantly differ

according to lender type: those that have lower warehousing costs are more likely to signal quality

than lenders with higher warehousing costs.

The signal is reflected in ex-post performance and initial offerings pricing, further strengthening

our interpretation of the results. A one standard deviation increase in signaling on soft information

4

lowers the odds ratio of charge-off in 3 months post-securitization by 41.3 percent, or in dollar

amounts, $1,562 less charged-off dollars from the principal loan amount. Moreover, signaling is

priced in initial offerings of the ABS securities. As our measure of signaling increases by one

standard deviation, the spread at origination on ABS over comparable-maturity Treasury securities

narrows by 0.15 percentage points. Of note, industry practice is to use the spread at origination

to indicate the underlying quality of the loans. Given the statistical significance of our signaling

measure, this suggests that the signal is used in part to overcome the lemons problem. That is,

while we do see that, in general, securities conform to the no-adverse-selection reps and warranties

requirement on information that is public, our private information results suggests a potential

adverse selection problem that is solved through the warehousing time signal.

Our paper makes two main contributions to the literature. First, we contribute to the literature

on asymmetric information in securitization markets. Pre-crisis work on securitization of bank loans

highlighted the role of internal funding costs, premiums associated with the sale of loans, and the

probability of bank default as key considerations regarding a bank’s decision to sell loans rather

than retain them on the balance sheet (Gorton and Pennacchi, 1995). Other papers highlighted

the possibility of efficient risk sharing and enhanced liquidity as another reason for securitization

(Merton, 1990). In the wake of the crisis, a host of papers illustrated the implications of asymmetric

information in securitization, primarily exploring the MBS market. For example, Keys et al. (2010)

demonstrate that the securitization process and asymmetric information led in part to lax screening.

Moreover, Rajan, Seru, and Vig (2015) show that unreported information can be more important

than reported information for predicting ex post performance of mortgages. In addition, Agarwal,

Chang, and Yavas (2012) illustrate that there was adverse selection in the securitization of prime

mortgages as a result of information asymmetries; Calem, Henderson, and Liles (2011) provide

evidence that this behavior resulted in “cherry-picking” by some mortgage lenders. However,

Albertazzi et al. (2015) discuss that, for reputation reasons, some mortgage lenders did not exploit

all information asymmetries; still, Griffin, Lowery, and Saretto (2014) provide evidence to the

contrary.

Against this backdrop, our contribution to the literature is to provide evidence of asymmetric

information and signaling in a post-crisis regulatory environment. Adelino, Gerardi, and Hartman-

Glaser (2019) illustrates that signaling was used to remedy informational asymmetries and the

5

signal they study is similar to ours – warehousing times. While their focus is on the pre-crisis

mortgage market, our paper examines the post-crisis auto ABS market, and we directly measure

asymmetric information instead of inferring it from ex-post performance. Neilson et al. (2019) show

the general effect of improved information via the new ABS-EE filings requirements post crisis. Our

contribution is that we use a further distinction of private vs. public information to identify the

effect of the post-crisis regulation on these different parts in the information structure. Moreover,

we exploit the requirement of “no adverse selection” in the construction of pools to gain further

insight into the effect of private information on signaling and pricing. More generally, our paper

demonstrates that, even with Dodd-Frank regulatory reforms, private information significantly

affects pricing, similar to Furfine (2019) for the CMBS market.

Second, our paper contributes to the literature on consumer ABS markets. Relative to the

literature on the MBS market, the literature on consumer ABS markets is thin. 4 Implications of

any findings on auto loan securitization could be potentially large, as roughly 85 percent of U.S.

households own some kind of vehicle, versus 64 percent who own a home (Bricker et al., 2017). In

addition, auto ABS provides an ideal laboratory to study asymmetric information because most

loans (in our data, all loans) are originated and securitized by the same entity and the loans are

highly standardized, compared to other types of loans including mortgages.

The remainder of the paper is organized as follows. Section 2 discusses the background and

the regulatory framework for auto ABS. Section 3 describes the data. Section 4 presents empirical

results from the baseline estimation for signaling, and two sub-sample robustness tests. Section 5

discusses three experiments and reports results. Section 6 presents results on post-securitization

auto loan performance. Section 7 discusses and tests for alternative hypotheses. Section 8 concludes.

2 Auto ABS: Background and Regulatory Framework

2.1 Background

Lending for automobiles has been one of the most commonly issued consumer credit loans, as

automakers find it advantageous to issue loans to boost sales. For example, The General Motors

4Faltin-Traeger, Johnson, and Mayer (2010) is an important exception regarding crisis-period pricing.42016 Survey of Consumer Finances.

6

Acceptance Corporation (GMAC) was founded in 1919 to provide credit for new-car buyers.5 Sim-

ilar to other consumer credit markets and the mortgage market, there was a buildup of lending

in the early-to-mid 2000s (Figure 1). At the same time, the rise was far less steep than that of

mortgages, and in particular, the market did not completely shut down during the financial crisis.

Today, auto loans are still one of the most commonly held forms of consumer credit: auto loans

have climbed to about $1.2 trillion outstanding at the end of 2018.

Auto ABS was one the first consumer ABS to come to the market in the 1980s, when securi-

tization of consumer loans grew in earnest. Auto ABS markets continued to function for most of

the financial crisis, although they received some support through the Term Asset-Backed Securities

Loan Facility (TALF) (Campbell et al., 2011). Moreover, some securitization markets shut down af-

ter the crisis — most notably, that for private-label mortgage-backed securities. However, the auto

ABS market has continued, even in riskier segments. At the end of 2018, auto ABS outstanding

stood near $225 billion, with issuance in that year around $107 billion.6

Auto ABS provides an ideal setting to study important economic questions for many reasons.

First, compared to other types of securities, auto ABS is a “plain vanilla” security. Just a few loan

characteristics — vehicle make and model, year, age, etc. — are enough to capture most variation

in collateral quality. On the other hand, in most other securitizations, it is extremely difficult to

summarize heterogeneous dimensions of collateral quality, such as those of commercial real estate

loans, with just a few numbers. Second, there has been little direct government intervention in

auto ABS outside of the financial crisis. It therefore allows a thorough look at the inherent market

mechanisms. Lastly, intermediation chains are short, thereby significantly reducing the burden to

consider other types of complexities caused by multiple intermediaries. For public deals, auto loan

originators are nearly always the servicer and the securitizer of the loan, and the sponsor of the

associated auto ABS.

Even though auto ABS are simple, most studies of securitization focus on the residential mort-

gage market; rightfully so, given its central role in the financial crisis, not to mention the $7 trillion

in agency MBS.7 Nevertheless, it is perhaps more challenging to explore issues related to adverse

5Henry Ford disapproved of providing credit, and provided a form of savings account instead (“The Engine ofEnterprise: Credit in America” by Rowena Olegario, Cambridge: Harvard University Press, 2016).

6Refer to SIFMA, U.S. ABS issuance and outstanding, available at https://www.sifma.org/resources/

research/us-abs-issuance-and-outstanding/.7Refer to SIFMA, U.S. Agency MBS issuance and outstanding, available at https://www.sifma.org/resources/

7

selection in securitization markets within the context of the mortgage market. It is not only com-

plicated by the existence of implicit or explicit government support from the government-sponsored

enterprises, but also, differences in collateral quality related to geographic region are substantial.

Finally, in the residential mortgage market, the loan originator, securitizer, servicer and sponsor

are often unrelated entities (Kim et al., 2018).

One complication of auto ABS relative to residential MBS to study asymmetric information

in securitization markets is that there can be changes in loan terms and securitization related

to auto demand, and these decisions are made in a very centralized way. That is, auto loan

terms, credit availability, and securitization practices are often adjusted to accommodate auto

demand or supply by the major automakers. Although there may be coordinated incentives at

selected steps in the intermediation chain, because the mortgage intermediation chain is more

segmented, such broad incentives are likely more limited in residential mortgage markets. That said,

this auto ABS centralization would tend to bias our results towards not finding significant effects

of private information, as changes in securitization practices from demand factors could swamp

those from private information. In addition, our securitizations represent better quality borrowers

and securities than those that comprise the privately-placed securitization market. Asymmetric

information is likely greater in the privately-placed market. Our results should be interpreted

against this backdrop accordingly.

2.2 Regulatory Framework

There are a few important regulatory characteristics about the auto ABS market that illustrate why

the auto ABS market is a powerful laboratory for exploring the effects of asymmetric information.

Specifically, as part of the Dodd-Frank Act, for all asset-backed securities transactions (auto ABS,

MBS and others), rating agencies must verify that all “representations (reps) and warranties” have

been satisfied.8 These reps and warranties confirm a number of attributes of the loan. Some

research/us-mortgage-related-issuance-and-outstanding.817 CFR Parts 232, 240, 249, and 249b states that:

As noted in Section 943 of the Dodd-Frank Act mandates that the Commission adopt rules requiring annationally-recognized statistical rating organization (NRSRO) to include in any report accompanying acredit rating of an asset-backed security a description of the representations, warranties, and enforce-ment mechanisms available to investors and how they differ from the representations, warranties, andenforcement mechanisms in issuances of similar securities.

See Pub. L. No. 111-203, 943.

8

examples of these are that the loan is a“valid sale and binding obligation,” the loan has scheduled

payments, the loan can be prepaid, and the obligor has auto insurance.

The presence of one rep and warranty and the absence of another deserve particular mention.

First, rating agencies must ensure that “no adverse selection” was used in the construction of

the underlying loan pool. Because of this, originators are expected to securitize (or not) a broad

spectrum of loans. Second, there is no rep and warranty for auto ABS dictating that issuers obtain

the same level of documentation that is required of other securitizations. For example, income

verification is required for mortgages underlying MBS, but not required for auto loans backing an

ABS. One possible reason for this is that overall, auto loan amounts are relatively low.

Taken together, these reps and warranties suggest a possible need to signal and a possible actual

signal in the auto ABS market. The adverse selection rule can present a problem to a lender that

has private information about the borrower. Specifically, the lender would lose money on a loan

that, according to observable characteristics, would command a lower price. In this situation, a

lender would want to signal the underlying quality of the loan. Moreover, the lack of a sufficient

documentation rule suggests that income verification or other information gathering is above and

beyond the basic requirements for the loans. If the lender does obtain extra information, it can use

this as a signal for the underlying quality of the loan. We use these characteristics in the analysis

that follows.

3 Data

Our primary data comes from loan-level XML files and prospectuses associated with post-crisis

reporting requirements under SEC Regulation AB. The reporting requirement went into effect

on November 23, 2016. Under the reporting requirement, all filings of prospectuses for public

securities offerings must also have accompanying loan-level information submitted in electronic

format, using SEC form ABS-EE.9 Prospectuses must be filed at least three days in advance of the

public placement of securities.10

The loan-level XML files comprise the universe of public auto ABS issued from 2017 Q1 to

9The requirement applies to all registered offerings backed by auto loans and leases, residential and commercialmortgages, and debt securities including re-securitizations.

10 More information is available at https://www.govinfo.gov/content/pkg/FR-2014-09-24/pdf/2014-21375.

pdf.

9

2019 Q1, and consists of 5,930,935 unique loans and 89 ABS securities issued by 15 issuers. The

data contains a range of information on the originator, borrower, and collateral associated with

each loan. Our main variable of interest, the loan warehousing time, is computed by measuring the

number of months between loan’s origination and securitization. Figure 2 shows the distribution

of warehousing time in our data. The average warehousing time is 15 months. While the majority

of loans in our sample is securitized within 10 months of origination, some securitized loans remain

on lenders’ books for up to five years.

Table 1 presents summary statistics of the remainder of auto loan characteristics from the loan-

level XML files, as of loan origination time. Because all ABS securities in our data are public

issuance, they mostly represent prime auto loans. An average borrower in our data has a credit

score of 695 and takes out a loan on a $25,427 (relatively new) car at 8% interest rate for 67 months.

Nevertheless, as can be seen from the wide ranges of borrower and collateral characteristics in Table

1, our data also has a substantial number of subprime auto loans. As such, the data allow us to

capture a broad view into securitization, across both prime borrowers and subprime borrowers,

used and new cars, and a range of geographic locations, not confined to specific groups of borrower

or collateral.

Figure 3 breaks down the data by origination year and originator type. Most originations in

our data are associated with captive auto finance companies and other non-bank finance companies

affiliated with auto retailers.11 Over the period, these non-banks altogether comprised roughly 70

percent of auto loan origination. Banks also have a significant presence representing the remaining

30 percent of origination, with some banks (e.g., Santander) lending primarily on the subprime

market and other banks servicing a variety of borrowers.12

While several proxies for soft information on the loans are available in the XML exhibits, they

are not included in any of the prospectuses for the deals. To exploit the difference in information

disclosure between the loan-level electronic exhibits and prospectuses, we match each ABS from the

XML files to the prospectuses filed to the SEC prior to the placement. All auto ABS prospectuses

10Original terms for the loans range from 24 months to 85 months in our data.10A large, remaining share of auto ABS is privately placed, and not offered on public markets. Privately-placed

auto ABS is disproportionately backed by subprime auto lending, and has been the focus of some attention lately.11Our data reflects nine captive auto finance companies: BMW, Daimler, Ford, GM, Honda, Hyundai, Nissan,

Toyota, and Volkswagen. Carmax, a used car dealer, also originates a notable share of loans.12Our data reflects five banks: Ally, California Republic, Fifth Third, Santander, and USAA.

10

have tear sheets that have summary statistics describing the composition of their collateral pools.

Figure 4 shows an example tear sheet from an actual ABS deal prospectus. As in this example, all

of the prospectuses matched to our data always report summary statistics on average warehousing

time, obligor credit score, payment-to-income ratio, original interest rate, original loan amount,

and original loan term. In contrast, none of them report summary statistics on any proxy for

soft information, such as the share of the pool that has been verified on income/employment or

made on under an exception. Because many ABS investors reportedly rely on deal prospectuses

to make investing decisions, information in the XML files is relatively more costly to obtain. This

distinction allows us refer to the first group of loan characteristics — that are disclosed both on

electronic exhibits and prospectuses — as variables indicative of “hard” information and the second

group of loan characteristics — that are only disclosed on the loan-level electronic exhibits and never

on prospectuses — as variables carrying “soft” information.

For auto recall-related analyses, we supplement the above loan-level SEC dataset with data

from the National Highway Traffic Safety Administration’s Office of Defects Investigation, IHS

Markit’s New Vehicle Registration, and Polk’s Vehicles in Operation. We also use Ravenpack News

Analytics to obtain sentiment and news volumes data on recalls (see the Data Appendix for more

details). For the analysis using model updates, we gather major model updates information on 44

makes during 2014-2018 from Kelley Blue Book. We use Bloomberg for ABS price analyses.

4 Empirical Results

4.1 Preliminaries

Figure 5 shows preliminary evidence of selective timing based on soft information. We divide the

sample into two groups based on the soft characteristics described above. In Figure 5a, we divide the

sample into loans made under an underwriting exception versus those made without any exception.

In general, lenders make loans that conform to set underwriting guidelines. However, whenever

necessary, lenders additionally obtain information about the collateral and the borrower, and make

underwriting exceptions upon seeing good mitigating factors to approve a loan. Therefore, the

loans approved under an exception have been verified of good soft information about the loan,

controlling for the usual loan characteristics. Similarly, in Figure 5b, the sample is divided into

11

loans made on new cars versus loans made on used cars.

Both sub-figures show a clear rightward shift in the empirical distribution of warehoused months

for loans that were likely of better credit or collateral quality, as opposed to other loans. Of course,

there could be other factors driving these results. In what follows, we perform a battery of tests to

determine whether soft information is driving these results. The results withstand the tests, and

confirm the intuition in the figures.

4.2 Baseline Estimation

Specification To examine the effects of various loan characteristics on a loan’s warehousing time,

we evaluate the following baseline specification on the cross-section of all observed loans:

wijtots = β0 + β1Sito + β2Hito + FEj + Y -QFEto + Y -QFEts + εijtots (1)

Our dependent variable is wijtots , defined as the number of warehoused months for loan i that was

originated at time to in state j and securitized at time ts.

Sito is a matrix of variables that proxy for soft, or private, information the lender has regarding

the quality of loan i originated at time to. We use three variables: “Income Verified,” indicating

that the underwriter verified the income of the borrower, “Underwriting Exception,” indicating

that the lender made an exception to its underwriting guidelines when extending the loan, and

“Used Car” indicating whether the loan’s underlying collateral is a used car.

We view our variables included in Sito as proxies for soft information for three reasons. First,

as mentioned earlier, while there are indicators of verification and exception in the more detailed,

loan-level electronic filings, summary statistics on these items are not reported on the associated

prospectuses. Prospectuses provide useful information about the deal to potential investors in

a convenient and accessible form; XML files require scouring and computing. For this reason,

information costs can be lower from the prospectuses than from the XML files.

Second, even if investors used the loan-level filings to assess the pool quality, the lender still has

private information on the actual verifications or exceptions made. Income verification is usually

done selectively on those loans whose obligor has a low credit score, a high loan-to-value, etc. That

12Indicators of used versus new cars are included in a subset of prospectuses.

12

the loan was extended tells us that there was confirmation of otherwise unobservable loan quality

upon income verification. Similarly, that an underwriting exception was made and that the loan

was ultimately extended implies presence of private information on the lender side — particularly,

one that was indicative of better loan quality.

Third, while new cars represent uniform collateral, used cars can have soft information regarding

quality. Although some lenders, such as banks, may not have private information on the quality of

the car, others, such as captive auto finance companies or Carmax, are likely to have such private

information.

In addition to our soft information measures, we include Hito , which is a matrix of variables

that proxy hard, or public, information on loan i originated at to. We include variables at the core

of underwriting standards, including the obligor credit score and the payment-to-income ratio. We

also incorporate information about the loan itself, including the original interest rate, amount, and

term. Of note, and unlike the variables in Sito , these hard information indicators are included in

all of the prospectuses associated with the ABS deals in our data. Moreover, these are the variables

on which, presumably, the reps and warranty of “no adverse selection” should be applied.

We include a number of other controls to eliminate potential confounding effects. For example,

we control for geographic location of the loan by the state in which the loan was originated (FEj),

which proxies for macroeconomic conditions in the local area or differences in lending conditions

or bankruptcy laws that could drive some of our results. In addition, we include fixed effects

for the origination year-quarter (Y -QFEto), and fixed effects on the year and quarter the loan

was securitized (Y -QFEts). Taken together, these factors control for possible secular trends in

securitization markets.

Additionally, we check robustness using two more groups of fixed effects. The first group uses

dummy variables indicating the vehicle manufacturer or the make to further control for unobservable

factors associated with the underlying collateral. Manufacturer/make information represents most

of the important differentiation across autos. The second group include dummy variables for the

originator type. These indicators capture costs of warehousing, which can differ significantly across

lender types. For example, banks generally have a low-cost deposit base that can be used to finance

auto loans, while a nonbank captive auto financing company needs to rely on more expensive and

risky wholesale funding. For all our specifications, we cluster standard errors at the ABS pool level.

13

Results Table 2 presents our baseline results. Each column presents results that include our Sito

and Hito matrices; the specifications differ by the types of control variables used.

Overall, the results strongly suggest signaling in auto ABS markets. As shown in the top portion

of the table, proxies of soft information carry significant effects on loan warehousing times in the

majority of specifications. In particular, if the lender verified the borrower’s income or made an

underwriting exception, warehousing months are substantially longer. Importantly, the magnitudes

are economically significant. Conditional on hard information on loan/borrower characteristics and

various fixed effects, the standard deviation of warehoused months is close to one month, at 1.040.

If income is verified, then the warehoused time increases by roughly another week, or 0.2 standard

deviations. If there is an underwriting exception, warehousing times stretch even further, to about

11 to 15 days longer, or 0.35-0.48 standard deviations.

Soft information on the quality of the underlying collateral also plays an important role. Even

after controlling for a range of characteristics, used cars experience shorter warehousing periods

and new cars longer ones. The effects are statistically significant for most specifications and the

economic magnitudes suggest that used cars shorten warehousing times by about 7 to 8 days,

or 0.24-0.28 standard deviations. While statistics on the share of used car loans in the pool are

available on some prospectuses, ABS issuers may still choose to signal in a simple way the quality

of the underlying portfolio, as used cars, by their very nature, are subject to private information.

Additionally, purchases of an older car may be a proxy for other soft information not observable to

the investor.

The bottom portion of the table displays estimated coefficients for factors associated with hard

information. In sharp contrast to our soft information results, proxies of hard information have

smaller effects on warehousing times. The two primary underwriting variables, obligor credit score

and payment-to-income ratio, do not show economically significant coefficients. All else equal, the

0.0004 coefficient on obligor credit score implies a 0.01 day fall in the number of warehousing days

for every standard deviation decrease in credit score (96 points). So, while statistically significant,

the effect is not economically meaningful. The coefficient on the payment-to-income ratio does

not have any statistical significance under all specifications. Other factors, such as the original

interest rate at which the loan was extended, the loan amount or the term, are not consistently

14

significant across specifications. Taken together, these suggest that the lender signals what it sees

as the underlying — especially, unobservable — better quality of the loan than what the observable

information indicates.

The results are robust across various specifications using different sets of fixed effects in columns

(1)-(5). When we add the securitization year-quarter fixed effects, we see some coefficients drop

in magnitude and others lose significance (column (1) versus column (2)). Most notably, we see

the economic magnitude of the credit score coefficient drop by tenfold, suggesting there are secular

trends in the credit score’s effect on warehousing times. At the same time, there is little difference

in the coefficients across specifications that control for vehicle manufacturer and make (columns

(3) and (4)). This suggests robustness to our specification as well as some uniformity of factors

affecting warehousing times by manufacturer. Looking broadly across columns (2) through (5),

coefficients are relatively constant across specifications, save the magnitude and significance of

income verification when controlling for lender type. We investigate the effects of different lender

types in more detail in Section 7.1.

4.3 Robustness: Sub-sample Analyses

In this section, we conduct additional analyses to check robustness of our baseline results across

different subgroups of auto loans and time periods.

New cars We first use the same baseline specification as in equation (1), but limit the sample

to loans extended on new cars. We do this to eliminate any potential effect from unobserved

factors related to collateral quality. Specifically, even after controlling for borrower characteristics

and vehicle make/model, used cars are relatively more heterogeneous in quality than new ones,

and therefore, in their collateral value. By focusing only on the sub-sample of new cars, we can

eliminate altogether the potentially remaining soft information on the underlying quality of the

collateral, and instead capture only soft information on the borrower.

Table 3 reports the results. Underwriting exceptions continue to lengthen warehousing times,

with only a little less economic significance than in the baseline results. Income verification keeps

its sign but loses statistical significance. Some of this result could reflect relatively more homo-

geneous new car buyer profiles with respect to income than for used car buyers, in that new car

15

buyers often have higher income and credit scores than used car buyers. However, there are still

important effects of underwriting exceptions. If there is an underwriting exception, warehousing

time lengthens by about 6 to 10 days, or 0.2-0.3 standard deviations. The results are statistically

significant at the 1 percent level. As in the baseline, the economic magnitude of the effect of soft

information on warehousing times swamps that of hard information. Coefficients on the obligor

credit score and payment lose significance, perhaps bolstering the hypothesis regarding homogene-

ity of borrowers for new cars. At the same time, the original interest rate, loan amount, and term

remain either statistically or economically insignificant.

New loans In another sub-sample analysis, we focus on loans originated after 2017, dropping loans

originated during 2010-2016. Benefits from using this post-2017 sub-sample are two-fold. First,

these loans were originated following the introduction of the new ABS-EE reporting requirement in

2016. Therefore, we can test whether our previous findings on signaling were mostly an artifact of

the pre-reform period. Second, the sub-sample allows us to suppress the effects of remaining time

fixed effects potentially correlated with our soft/hard information variables and the warehousing

time. Because our main loan-level data comes from the reportings by ABS issuers since the 2016

reform, we only observe the loans securitized after 2017. By construction, warehousing time is

mechanically higher for those loans that we observe to be originated in earlier years. If there is a

remaining time trend in our soft/hard information variables (even after controlling for the year-

quarter fixed effects as in the baseline), then it can confound identification of the effects of our

soft/hard information variables on warehousing time.

Table 4 reports results based on the sample of post-2017 originated loans. In some cases,

economic magnitudes of the effects of soft information variables are larger than those in the baseline

results; in others, the results are roughly unchanged from the baseline. If income is verified, then

the warehouse time increases by 8 to 9 days, or around 0.3 standard deviations. This effect is 1 to 2

days longer than the baseline effect, suggesting some strengthening of the signal with the advent of

the reporting requirement. If a loan was extended under an underwriting exception, warehousing

time increases to about 14 to 19 days longer, or 0.44-0.62 standard deviations. The effects, again,

are larger in economic magnitude than those from the baseline effects (11-15 days). Coefficients on

the indicator for a used car loan are similar to those in the baseline. Used car loans are warehoused

16

for an average of 7 to 8 days, or around 0.2 standard deviations shorter than for new car loans.

Taken together, the results from the post-2017 originated loans sub-sample confirm our previous

findings, and suggest that our results identify signaling of private information that remains even

after the new disclosure requirement. More generally, these results highlight important information

asymmetries inherent in auto ABS. Against this backdrop, we turn to further tests of private

information below.

5 Evidence from Three Experiments

Other explanations may still be possible for our findings on signaling. For instance, there could

be remaining unobserved soft information about the quality of each loan that may be correlated

with the decision on the warehousing period. If the information is correlated with our variables of

soft or hard information, an endogeneity issue would arise. Furthermore, the true hazard rate of

securitization may not have been well captured because our data is only based on securitized loans.

It could also be the case that the longer warehousing times are, more fundamentally, not a signal

of private information. In this section, we bolster our previous findings on asymmetric information

and signaling by using three variables that are plausibly exogenous, but can affect the warehousing

period. These variables represent private information to lenders that they may want to signal to

investors, so that it will be appropriately priced in the market.

5.1 Variations in Collateral Quality: Auto Recalls

Our first experiment uses instances of auto recalls to explore the effects of exogenous variation in

the value of the collateral. As mentioned earlier, one strength of our analysis, relative to that on

mortgages, is that the quality of cars is reasonably uniform for a given make and model of car.

However, auto recalls represent an unexpected arrival of soft information regarding the underlying

collateral quality. So that the loans associated with non-recalled autos (presumably of higher

quality) are not priced and lumped together with loans for recalled autos (presumably lower quality),

lenders signal the higher quality collateral by warehousing for longer times.

To explore the effect of recalls on warehousing times, we collect auto recalls data from the

National Highway Traffic Safety Administration’s Office of Defects Investigation. We extract data

17

on all instances of auto recalls that took place between 2009 and 2018 for the automakers in our

data. Table 5 summarizes 1,601 recalls that took place during the period. A large majority are

trivial recalls, involving, say, wiper blades or headlights, met only with slight annoyance by the car

owner and handled with form letters and trips to the dealer. But some recalls involve more serious,

safety-threatening issues that make the front pages for months.

To account for differential severity of recalls, we combine the actual recalls information with

news reports on the recalls. Specifically, we use data on news volume and news sentiment on each

recall, from “Aggregate Event Volume” and “Aggregate Event Sentiment” from Ravenpack News

Analytics. Each of these data items measures the volume of events reported in financial news

and the share of positive sentiment events, on a 91-day rolling basis (see Data Appendix for more

details). The table shows wide ranges of average news volume, from 8.07 (Subaru) to 1741.33 (GM),

and average sentiment, from the lowest (i.e., the worst sentiment) 48.25 (FCA) to the highest (i.e.,

the best sentiment) 61.42 (Mclaren).

We also take into account the number of affected cars relative to the total number of cars on

the road. Because automakers have different market shares, some automakers will have many more

cars than others. Therefore, even if the affected units amount to 0.1-0.2 million vehicles, it does

not necessarily imply that the recall had a higher impact than a recall with a few hundreds or

thousands of affected units; it depends on the automaker. Using data from IHS Markit’s New

Vehicle Registration and Polk’s Vehicles in Operation, we scale the affected units of each recall by

the number of operating cars of the make that is impacted by the recall (“scaled affected units”).

In the end, we define Recall Impactkt as the (logarithm of) impact of recalls by an automaker

k in month t, taking into consideration the severity of the recall using the adjustments described

above.

Recall Impactkt = ln

(ScaledAffectedUnitskt ∗NewsV olumekt

Sentimentkt

)(2)

Scaled Affected Unitskt is the total scaled affected units of k in month t. News Volumekt and

Sentimentkt each takes an average of the news volume and the sentiment, for all recalls that took

place to k in each month t. In other words, the average recall impact for an automaker k is higher,

the higher the scaled affected units and the news volume, and the lower (or, the worse) the news

sentiment. Overall, this recall impact measure seems to be consistent with heuristic evidence from

18

well-known recalls. Figure 6 shows the Recall Impactkt of three automakers that experienced well-

publicized recalls: Toyota (gas pedal recalls of 2009-2011), GM (ignition switch recalls of 2014), and

Volkswagen (emissions recalls of 2016). For example, the GM ignition switch issue was primarily

focused on older cars, and so it should have less of an impact on sales according to our measure.

By contrast, the Toyota gas pedal recall implicated newer cars, and so that characteristic, coupled

with intense news coverage, causes the Toyota index to jump substantially more than does the GM

recall in 2014.

Finally, we define our main variable, 1-MoPost-Orig Recall Impactit, for loan i in month t as

the recall impact of i’s automaker k in month t+ 1, that is one month after loan i’s origination.

1-MoPost-Orig Recall Impactit = Recall Impactik,t+1 (3)

Notice that, to strengthen the “surprise” element, we use the recall impact one month after the

loan origination in our specification. By doing so, we eliminate any possible contamination of the

coefficient from demand or other effects resulting from the recall. As in the baseline estimation, we

keep our original proxies for soft and hard information.

We run our baseline equation, now adding this recall impact index. Table 6 Panel A presents the

results. In Column (1), we run the regression on all loans save those extended by captive auto finance

companies and, in Column (2), we run the same regression only on captive auto loans. Consistent

with our hypothesis, Column (1) shows that warehousing times shorten for loans associated with

recalls. In terms of economic magnitude, a 10% increase in the 1-month post origination recall

impact results in around 7 days shorter warehousing time. The result is statistically significant at

the 5 percent level.

Interestingly, in line with our signaling hypothesis, we see no differential effect on warehousing

times when we run the same regression on loans whose lenders were captive auto finance companies,

in Column (2). The absolute majority of auto loans extended by captive auto finance companies are

made on their own cars. By contrast, other types of lenders extend loans on various makes of cars.

Because investors are aware of this, they will immediately expect an ABS security issued by the

recall-implicated automaker to include loans extended on the affected cars. But, in ABS securities

issued by non-captive auto lenders, investors will be less certain regarding the extent of an auto

19

recall’s impact on the collateral pool. As such, captive auto lenders do not need to signal, while

other types of lenders do. In our regressions, this difference manifests as a significant coefficient on

our recall impact variable only for the sample of non-captive auto loans.

5.2 Variations in Borrower Quality: “Super Sales”

Our second experiment uses “super sales,” or periods of high incentives for car sales. In these

super sales periods, lenders may be more willing to extend loans to unobservably worse borrowers

in order to meet sales targets. Importantly, origination months are not summarized on auto ABS

prospectuses. As a result, this information is both costly to obtain for inattentive investors and

can proxy for private information. Our hypothesis is, then, that the month of origination could

proxy for soft information the lender has about the borrower’s credit quality.

To define periods of super sales, We identify months of the year that usually have substantial

rebates and incentives for car purchases: Memorial Day (May), Black Friday (November), and

end-of-year (December). We create two dummy variables, “Holiday” and “End-of-year”. These

indicators are equal to 1 if the loan was extended either in May or November, and in December,

respectively, and 0 otherwise.

Table 6 Panel B summarizes the results. The results suggest that warehouse times are sig-

nificantly shorter for loans extended during the super sale periods. For end-of-year origination,

warehousing times are trimmed by nearly 50 days. Loans extended during holiday sales also ex-

perience nearly two weeks shorter warehousing periods than the average loan. Both effects are

also statistically significant at the 1 percent level. Importantly, these findings are conditional on

the key underwriting variables including the primary underwriting variables of credit score and

payment-to-income ratio. It implies that during periods of heightened desire for sales, lenders may

provide funds for borrowers who, while observationally equivalent to other borrowers in hard char-

acteristics, would not necessarily get a loan during normal times. Consequently, for the market to

price the securities accurately, lenders would need to signal the quality of their collateral.

5.3 Variations in Borrower Quality: Model Changes

Similar to the second experiment, our final experiment uses another instance when there could

be extraordinarily heightened incentives for car sales — when there are major model changes.

20

To clear inventory of older models in advance of the introduction of new ones, automakers often

take aggressive marketing tactics to sell older models. Our hypothesis is that, similar to the

second experiment, lenders may be more willing to extend loans to unobservably worse borrowers

in advance of the new model introduction.

To investigate this possibility, we gather 240 major model updates on 44 makes that took place

during 2014-2018, from Kelley Blue Book.13 Because data on exact start dates of the new model

sale is not available (we only know the years of the major model updates), we approximate these

dates by taking the earliest date of loan origination on the corresponding model that we observe

in our data. We then define a dummy variable, “Model Update,” that is equal to 1 if the loan

was originated prior to a model update associated with the car, and 0 otherwise. Because our new

model sale dates are estimates, we explore different windows of 1, 3, and 6 months prior to the first

sale of each new model.

Table 7 displays the results. Conditional on our baseline specification plus the control for model

update, if a loan was originated one month prior to a model update, the loan’s warehousing time

decreased by roughly 4 days, or 0.13 standard deviations. The effect is statistically significant at

the 1 percent level. Loans originated significantly before a model update, for instance, 3 months

and 6 months before, do show shorter warehousing times in Column (2) and (3), but the effects

are not statistically significant. More generally, our findings imply that the heightened incentive

for older model sales seems to be strongest closer to the immediate month of the new model sale.

6 Post-Securitization Loan Performance

Our baseline results found evidence on asymmetric information and signaling based on soft infor-

mation observed by the lenders. The premise was that the proxies we used for soft information are

indicative of some positive or negative borrower/collateral quality. In this section, we check the

validity of our premise by examining if our ex ante proxies for soft information on borrower and

collateral quality are consistent with ex post loan outcomes.

We start by measuring a pool-k securitized loan i’s signaling of private information, Sikto , as

13To the best of our knowledge, these updates represent “major” model updates that generally occur every threeto five years, not minor updates that occur annually.

21

follows:

Sikto = β11 × IncomeV erifiedikto + β12 × Exceptionikto + β13 × UsedCarikto (4)

We use the estimated coefficients, β11, β12, and β13, from the baseline estimation in Table 2 Column

(3). In a nutshell, a higher Sikto implies signaling of positive soft information on loan i (shown

through the lender’s longer warehousing). On the contrary, loans with low Sikto indicate that those

loans could be of unobservably worse quality by, for instance, not being verified of income or not

having any exceptionally good credit factors, or by being extended on an used car.

Using the full panel data tracking each loan i securitized in pool k, we examine the relation-

ship between the loan’s summary statistic on signaling, Sikto , with its post-securitization principal

charge-off. In general, if a borrower remains delinquent after extensive collection efforts, lenders net

the sale proceeds from the repossessed vehicle, etc., from the remaining loan amount and charge off

the remaining balance owed by the borrower. Afterwards, the loan is usually replaced with another

loan of a similar risk profile. Our measure of charge-off is, in particular, useful in learning about the

intensive margins of loan performance because it continuously captures the dollar amount charged

off, in addition to the extensive margins such as becoming a delinquent account. Furthermore, the

measure also tracks the amount of the principal that is charged off due to loan modification.

Table 8 Panel A reports results from estimating a logit model of post-securitization probability of

principal charge-off and Panel B reports those from running an OLS regression of the dollar amount

of principal charge-off on Sikto . The outcomes show that our summary statistic for signaling is highly

associated with the loan’s probability of being charged off at 3 months of post-securitization and

the associated amount is substantial. A one standard deviation increase in Sikto lowers the odds

ratio of charge-off in 3 months by 41.3%, or in dollar amounts, $1,562 less charged-off dollars from

the principal loan amount. Since the average loan in our sample is $23,978, it implies that the

economic magnitude is tantamount to nearly 7 percent of the original loan amount. The effects

are similar and even tend to slightly increase in terms of charge-off probability when charge-off is

defined based on longer windows of time since securitization. All effects are statistically significant

at the 1 percent level.

Outcomes on hard loan characteristics are generally in line with our expectations, and are

22

economically meaningful. Loans with higher original interest rate, higher payment-to-income ratio,

lower original loan amount, and/or higher term have higher probabilities of principal charge-off. A

one standard deviation increase in credit score (96 points), payment-to-income ratio (4%), original

interest rate (7%), and loan term (8 months) raise the odds ratio of loan’s principal charge-off by

17.8, 8.9, 64.2, and 20.6 percent, respectively. A one standard deviation increase in original loan

amount ($10,255) decreases the odds ratio by 18.1 percent.

7 Alternative Explanations

While we believe that our results are consistent with signaling in the auto ABS market, there may

be alternative explanations. For example, our results could simply represent a “market for lemons”:

the bad loans, even though they are securitized faster, are priced the same as the better loans. This

would be a factor if the information content implicit in the loans is not priced. In this section,

we take two strategies to rule out these competing hypotheses. The first explores differences in

warehousing times according to originator type. Originators face different funding costs—banks

have deposit bases and so likely have lower funding costs, while non-banks and captive automakers

rely on wholesale short-term funding. The second investigates pricing implications on the ABS

pools. Both sets of results seem to point to signaling.

7.1 Funding Costs

As discussed in Section 3, there are different types of auto loan originators. One key distinguishing

feature of these lenders is their funding costs. Because banks have a deposit base, their funding

costs are usually considered to be lower than those of non-banks. On the other hand, non-bank

lenders rely more heavily on short-term wholesale funding and debt financing, compared to banks.

These alternative funding sources are often expensive, and in case of short-term funding, flighty.

Therefore, this difference in funding costs across lender types may result in a difference in the

average warehouse time of loans.

Table 9 indeed shows that warehousing times are longer for bank loans in general, compared

to loans extended by captive auto finance companies and other non-bank auto finance companies.

While the overall average warehousing time is 15 months for the whole sample, loans made by

23

captive autos have one month longer warehousing, at 16 months. Banks follow with 13 months

of average warehousing. However, there is an important difference between warehousing times for

Santander and other banks. In particular, Santander tends to focus on the subprime market, which

has lower quality borrowers. Splitting the bank sample into two groups—one without Santander,

and Santander itself—reveals that banks excluding Santander have the longest warehousing times on

average. Non-bank finance companies have a shorter-than-average warehousing time of 12 months.

As discussed in (Leland and Pyle, 1977), the decision to self-invest by lengthening warehousing

times depends on the relative risk pricing of good versus bad loans in the market, versus the costs of

warehousing. If lender warehousing costs are important, then banks should be able to signal better

loans through longer warehousing times than the non-banks. In particular, some nonbanks likely

obtain funding from warehouse lenders. These lenders sometimes require that loans are securitized

within a specific window. As such, it may be more expensive for nonbank lenders to fund loans for

a longer time than banks do.

To evaluate this possibility, we explore the baseline specification in equation (1), but limit our

samples to one lender type only. Table 10 presents the results. Columns (1)-(3) each report results

from running the regression on the sub-samples of loans extended by banks, captive autos, and

other non-banks. Looking across the lender types, all else equal, and for any one element of soft

information, banks appear to shift warehousing times based on soft information most significantly.

Banks warehouse loans whose obligor income has been verified for roughly 4 days longer, or 0.12

standard deviations. The effect is about half of that of the baseline result but is statistically sig-

nificant at 1% level. Neither captive autos nor other non-bank finance companies has a significant

coefficient on the indicator for income verified loans. Furthermore, for any single instance of un-

derwriting exception, banks warehouse loans for around 17 more days, or 0.55 standard deviations.

The economic magnitude is five times as long as that of captive autos. Loans made on used car have

the least effect, with around 2 more days in banks’ warehousing time. Again, non-bank lenders do

not show any significant effects of used car loans on warehousing period. Our results suggest that

warehousing times are consistent with funding cost differentials, further bolstering our signaling

argument.

24

7.2 Pricing Implications

Another competing hypothesis is that our results capture only adverse selection, and not signaling.

If this were the case, then we would assume investors to view the market as one for “lemons,” where

pooling occurs at an equilibrium in which only the lower-quality loans are sold, and the better ones

are retained.

To investigate this possibility, we turn to an analysis of ABS pricing. We aggregate our loan-

level data to the ABS pool-level. We define a pool k-level summary statistic of signaling based on

private information by taking a weighted average of Sikto of all loans securitized in pool k, where

Sikto is as defined in Equation (4):

Skt =∑i∈k

ωiktSikto

=∑i∈k

ωikt × (β11 × IncomeV erifiedikto + β12 × Exceptionikto + β13 × UsedCarikto)

(5)

ωikt is the origination loan amount share of loan i in pool k. A higher Skt implies that the pool k

contains a higher share of unobservably good quality loans — in form of loans that are verified of

income, that have exceptionally good credit factors, or that are made on new cars.

Table 11 describes the pool-level summary statistics of 89 auto ABS securities included in our

sample. Skt exhibits a high variation across the pools. While the average pool in our data has

near zero change in warehousing time arising from soft information, the standard deviation of

warehousing time change due to soft information is around 4 days. Among all pools, the signaling

of the highest quality loans results in up to seven days longer warehousing times; the lowest quality

loans experience up to a nine days shorter warehousing time. The average warehouse time of loans

also greatly varies from 2 months to around 2 years across the pools. In addition to soft information,

hard information can be associated with wider or narrower ABS spreads over comparable maturity

riskless securities. Therefore, we also take into account other pool-level characteristics such as

average maturity and credit score of the loans in the pool.

Based on the pool-level characteristics and our summary measure of signaling, δ1Skt, we evaluate

25

the following specification:

rkt = δ0 + δ1Skt + δ2Xkt + γt + µkt. (6)

Our dependent variable rkt is the average yield spread of pool k issued at time t over comparable

5-year Treasury securities. Xkt is a vector of pool k characteristics at time t, including k’s face

amount and number of tranches, and the number and the average credit score of the loans included

in k. γt is issuance time fixed effects. µkt is an error term. We control for the year and quarter

of securitization to account for secular trends in the demand for auto ABS securities. We also

control for lender type fixed effects to control for different demand and appetite for certain type’s

securitization (for instance, Santander’s sub-prime auto ABS). We bootstrap the standard errors

to control for understatement of the standard errors as a result of including our summary measure

for signaling, ckt, as one of the right-hand side variables.

Table 12 presents our pricing analysis results. Consistent with our signaling hypothesis, we find

that soft information is priced. In Column (1), if Skt increases, all else equal, spreads on ABS fall.

The effect is statistically significant at the 5% level. Economic magnitude is meaningful as well; a

one standard deviation increase in Skt is associated with a 0.15 percentage point, or a 0.39 standard

deviations, decrease in the ABS spread.

In addition to Skt — the weighted average of Sikto of all loans in pool k — in Column (1),

we explore other types of summary statistic for Sikto in Columns (2)-(4), in terms of median,

minimum, and standard deviation of Sikto . Results are similar for the median and the minimum

levels of signaling for pool k collateral. As the median and the minimum Sikto in pool k increase by

one standard deviation, ABS spreads decrease by 0.08 and 0.03 percentage points respectively, or

0.23 and 0.08 standard deviations. The effect in terms of minimum Sikto is statistically significant

at the 5% level. The strongest economic effect is for the standard deviation of loan-level signaling

in Column (4). As the pool’s standard deviation of Sikto increases by one (across-pool) standard

deviation, the spread decreases by 0.24 percentage points, or 0.63 standard deviations. The result

is consistent with self-financing signaling. Even if there is still some pooling in the pricing of the

ABS, the existence and signaling of the higher-quality loans leads investors to view the overall

market as favorable.

Outside of the private information, we find that various measures of public information are also

26

priced, in expected ways. Larger loans and higher credit scores tend to receive favorable pricing,

though the size of the loan lacks in statistical significance. Results on number of tranches are neither

economically nor statistically significant. Excessive tranching can indicate less self-financing and

results in wider spreads.

8 Conclusion

In this paper, we leverage two different datasets with varying levels of information disclosure to

define soft and hard information on auto loans. Such distinction allows for a direct measurement

of asymmetric information and exploration of the effects. Our results suggest that asymmetric

information is alive and well in the auto ABS market, despite the new post-crisis reporting require-

ments. While most previous studies find asymmetric information in residential mortgage markets,

this paper underpins the inherent nature of asymmetric information in, more broadly, securitization

by finding such evidence based on the data that only includes publicly-placed auto ABS, which are

considered to comprise the most highly-rated securities.

Overall, this work brings to the forefront the question of whether securitization is beneficial

risk-sharing, or if it is inherently risky money creation. Reporting requirements can eliminate

some, but not all, risky aspects of securitization. As such, securitization is inherently risky, and

should continue to be monitored. Our results can inform other regulatory initiatives surrounding

securitization.

27

References

Adelino, Manuel, Kristopher Gerardi, and Barney Hartman-Glaser. 2019. “Are lemons sold first?

Dynamic signaling in the mortgage market.” Journal of Financial Economics 132 (1):1 – 25.

URL http://www.sciencedirect.com/science/article/pii/S0304405X18302575.

Agarwal, Sumit, Yan Chang, and Abdullah Yavas. 2012. “Adverse selection in mortgage securiti-

zation.” Journal of Financial Economics 105 (3):640–660. URL https://EconPapers.repec.

org/RePEc:eee:jfinec:v:105:y:2012:i:3:p:640-660.

Akerlof, George A. 1970. “The Market for “Lemons”: Quality Uncertainty and the Market Mech-

anism*.” The Quarterly Journal of Economics 84 (3):488–500.

Albertazzi, Ugo, Ginette Eramo, Leonardo Gambacorta, and Carmelo Salleo. 2015. “Asymmetric

information in securitization: An empirical assessment.” Journal of Monetary Economics 71:33

– 49. URL http://www.sciencedirect.com/science/article/pii/S0304393214001597.

Bricker, Jesse, Lisa J. Dettling, Alice Henriques, Joanne W. Hsu, Lindsay Jacobs, Kevin B. Moore,

Sarah Pack, John Sabelhaus, Jeffrey Thompson, and Richard A. Windle. 2017. Federal Reserve

Bulletin 103 (3):1–42.

Calem, Paul, Christopher Henderson, and Jonathan Liles. 2011. ““Cherry picking” in subprime

mortgage securitizations: Which subprime mortgage loans were sold by depository institutions

prior to the crisis of 2007?” Journal of Housing Economics 20 (2):120 – 140. URL http://www.

sciencedirect.com/science/article/pii/S1051137711000167. Special Issue: Housing and

the Credit Crunch.

Campbell, Sean, Daniel Covitz, William Nelson, and Karen Pence. 2011. “Securitization markets

and central banking: An evaluation of the term asset-backed securities loan facility.” Jour-

nal of Monetary Economics 58 (5):518 – 531. URL http://www.sciencedirect.com/science/

article/pii/S0304393211000468. Carnegie-Rochester Conference on public policy: Normaliz-

ing Central Bank Practice in Light of the credit Turmoil, 12–13 November 2010.

Downing, Chris, Dwight Jaffee, and Nancy Wallace. 2008. “Is the Market for Mortgage-Backed

Securities a Market for Lemons?” The Review of Financial Studies 22 (7):2457–2494. URL

https://doi.org/10.1093/rfs/hhn114.

Faltin-Traeger, Oliver, Kathleen W. Johnson, and Christopher Mayer. 2010. “Issuer Credit Quality

and the Price of Asset Backed Securities.” American Economic Review 100 (2):501–05. URL

http://www.aeaweb.org/articles?id=10.1257/aer.100.2.501.

Furfine, Craig. 2019. “The Impact of Risk Retention Regulation on the Underwriting of Securitized

Mortgages.” Journal of Financial Services Research .

28

Gorton, Gary and Andrew Metrick. 2013. “Securitization.” chap. Chapter 1. Elsevier, 1–70. URL

https://EconPapers.repec.org/RePEc:eee:finchp:2-a-1-70.

Gorton, Gary B. and George G. Pennacchi. 1995. “Banks and loan sales: Marketing nonmarketable

assets.” Journal of Monetary Economics 35 (3):389 – 411. URL http://www.sciencedirect.

com/science/article/pii/030439329501199X.

Griffin, John, Richard Lowery, and Alessio Saretto. 2014. “Complex Securities and Underwriter

Reputation: Do Reputable Underwriters Produce Better Securities?” The Review of Financial

Studies 27 (10):2872–2925. URL http://www.jstor.org/stable/24466857.

Keys, Benjamin J., Tanmoy Mukherjee, Amit Seru, and Vikrant Vig. 2010. “Did Securitization

Lead to Lax Screening? Evidence from Subprime Loans.” The Quarterly Journal of Economics

125 (1):307–362. URL https://doi.org/10.1162/qjec.2010.125.1.307.

Keys, Benjamin J., Amit Seru, and Vikrant Vig. 2012. “Lender Screening and the Role of Secu-

ritization: Evidence from Prime and Subprime Mortgage Markets.” The Review of Financial

Studies 25 (7):2071–2108. URL http://www.jstor.org/stable/23263559.

Kim, You Suk, Steven Laufer, Karen Pence, Richard Stanton, and Nancy Wallace. 2018. “Liquidity

crisis in the mortgage market.” Brookings Papers on Economic Activity .

Leland, Hayne and David H Pyle. 1977. “Informational Asymmetries, Financial Structure, and Fi-

nancial Intermediation.” Journal of Finance 32 (2):371–87. URL https://EconPapers.repec.

org/RePEc:bla:jfinan:v:32:y:1977:i:2:p:371-87.

Merton, Robert C. 1990. “The Financial System and Economic Performance.” Journal of Financial

Services Research 4 (4):263–300.

Neilson, Jed J., Stephen G. Ryan, K. Philip Wang, and Biqin Xie. 2019. “Asset-level Transparency

and the (E)valuation of Asset-Backed Securities.” Working Paper 20777, SSRN. URL https:

//ssrn.com/abstract=3247342.

Rajan, Uday, Amit Seru, and Vikrant Vig. 2015. “The failure of models that predict failure:

Distance, incentives, and defaults.” Journal of Financial Economics 115 (2):237–260. URL

https://EconPapers.repec.org/RePEc:eee:jfinec:v:115:y:2015:i:2:p:237-260.

Spence, Michael. 1973. “Job Market Signaling.” The Quarterly Journal of Economics 87 (3):355–

374. URL http://www.jstor.org/stable/1882010.

29

Figures and Tables

Figure 1: Auto Loans Outstanding

Source: Federal Reserve Board Statistical Release G.19, Consumer Credit, Automobile Loans.

30

Figure 2: Distribution of Warehoused Months

Note: This figure displays the distribution of warehoused months of loans in our dataset. The dataset comprises theuniverse of public auto ABS issued from 2017 Q1 to 2019 Q1, consisting of 5,930,935 loans and 89 ABS securitiespublicly issued by 15 issuers. Warehoused months are winsorized at bottom and top 1%. Source: SEC form ABS-EE.

31

Figure 3: Number of Loans by Origination Year and Lender Type, 2010-2018

Note: This bar chart shows the total number of loans originated in each year from 2010 to 2018 by each lender type:captive auto finance companies (BMW, Daimler, Ford, GM, Honda, Hyundai, Nissan, Toyota, and Volkswagen),other non-bank finance companies (Carmax), banks exluding Santander (Ally, California Republic, Fifth Third, andUSAA), and Santander.

32

Figure 4: Description of Pool Composition on Auto ABS Prospectuses: An Example

Note: This example, from Ford Credit Auto Owner Trust 2018-A, shows a page from an auto ABS prospectus wherethe issuer discloses summary statistics of various characteristics on their collateral pool. Source: SEC EDGAR.

33

Figure 5: Warehousing of Loans When Some Information is Unavailable on Deal Prospectuses

(a) “Exception” vs. “Normal” Loans

(b) New Car vs. Used Car Loans

Note: This figure shows a shift in the empirical distribution of warehoused months in an instance when only thelender knows that a loan is of good quality, i.e., when the lender has private information that the loan is good.Subfigure (a) compares loans that were of exceptionally good quality and therefore extended under an underwritingexception (“Exception” loans), to loans extended without any exception (“Normal” loans). Subfigure (b) comparesloans made on new cars against those made on used cars. The sample comprises of 5,930,935 loans collateralized in89 public auto ABS issued by 15 issuers from 2017 Q1 to 2019 Q1. Warehoused months are winsorized at bottomand top 1%. Source: SEC form ABS-EE.

34

Fig

ure

6:R

ecal

lIm

pac

tsof

Sel

ect

Au

toM

anu

fact

ure

rs,

2009

-201

8

Fig

ure

7:

Note

:T

hes

efigure

ssh

owth

e”R

ecall

Impact

”va

riable

for

Toy

ota

,G

ener

al

Moto

rs,

and

Volk

swagen

(“i”

)duri

ng

2009-2

018

(“t”

,m

onth

ly),

and

captu

res

the

noto

rious

reca

llin

ciden

tsin

cludin

gT

oyota

gas

ped

al(2

009-2

011),

GM

ignit

ion

swit

ch(2

014),

and

Volk

swagen

die

selem

issi

ons

(2016).

Rec

all

Impact

isdefi

ned

asRecallImpact

kt

=ln

(ScaledAffectedUnits k

t∗NewsVolume k

t/Sentiment k

t)

for

an

auto

makerk

inm

ontht.ScaledAffectedUnits k

tadju

sts

the

impact

of

each

reca

llby

scaling

the

aff

ecte

dunit

sof

reca

lled

veh

icle

sby

the

esti

mate

dnum

ber

of

the

veh

icle

sof

reca

lled

model

-yea

rth

at

rem

ain

on

the

road

at

the

tim

eof

the

reca

ll.NewsVolume k

tandSentiment k

tea

chco

mes

from

Rav

enpack

New

sA

naly

tics

’“A

ggre

gate

Even

tV

olu

me”

and

“A

ggre

gate

Even

tSen

tim

ent,

”co

ndit

ional

on

the

rele

vance

of

the

new

sev

ent

bei

ng

larg

erth

an

75

(on

a1-1

00

scale

);ea

chm

easu

reth

evolu

me

of

even

tsre

port

edin

financi

al

new

sm

ediu

mand

the

share

of

posi

tive

even

ts,

on

a91-d

ayro

llin

gbasi

s.Sourc

e:N

ati

onal

Hig

hw

ayT

raffi

cSafe

tyA

dm

inis

trati

ons

Offi

ceof

Def

ects

Inves

tigati

on,

IHS

Mark

it’s

New

Veh

icle

Reg

istr

ati

on,

Polk

’sV

ehic

les

inO

per

ati

on,

and

Rav

enpack

New

sA

naly

tics

.

35

Table 1: Summary Statistics of Borrower and Collateral Characteristics at Origination

Mean Stdev Min Max

Original Interest Rate 0.08 0.07 0.01 0.27Original Loan Amount ($) 23,978 10,255 4,500 130,485Original Loan Term (Months) 67 8 24 85Obligor Credit Score 695 96 437 847Vehicle Age (Years) 1.4 2.0 0.0 14.0Used Car 0.43 0.49 0 1Vehicle Value Amount ($) 25,427 11,524 4,539 135,310Payment-to-Income Ratio 0.08 0.04 0.01 0.28Income Verified 0.08 0.27 0.00 1.00Employment Status Verified 0.07 0.25 0.00 1.00Underwriting Exception 0.18 0.38 0.00 1.00End-of-Year Origination 0.08 0.27 0.00 1.00Holiday-Sale Origination 0.17 0.37 0.00 1.00

Note: This table presents summary statistics for our main variables. Our data comprises the universe of public autoABS issued from 2017 Q1 to 2019 Q1, and consists of 5,930,935 loans and 89 ABS securities publicly placed by 15issuers. “Underwriting Exception” indicates if the loan was an exception to underwriting criteria. “End-of-YearOrigination” indicates if the loan was originated at the end of a year (defined here as in December). ”Holiday-SaleOrigination” indicates if the loan was originated in the months of May (Memorial Day sale) or November (BlackFriday). ”Used Car,” “Income Verified,” “Employment Status Verified,” “Underwriting Exception,” “End-of-YearOrigination,” and “Holiday Sales” each indicate a share of loans as a percentage satisfying each of the categoricalconditions. Loans with “Original Interest Rate” over 100% or “Payment-to-Income Ratio” over 0.5 are dropped.“Original Interest Rate,” “Original Loan Amount,” “Original Loan Term,” “Obligor Credit Score,” “Vehicle Age,”“Vehicle Value Amount,” and “Payment-to-Income Ratio” are winsorized at each originator’s bottom and top 1%.Source: SEC form ABS-EE.

36

Table 2: Asymmetric Information on Borrower and Collateral Characteristics

Dependent Variable: Warehoused Months(1) (2) (3) (4) (5)

“Soft” Information

Income Verified 0.311 0.209** 0.223** 0.232*** 0.0876(0.382) (0.095) (0.093) (0.093) (0.094)

Underwriting Exception 2.171*** 0.496*** 0.416*** 0.397** 0.360***(0.681) (0.099) (0.100) (0.096) (0.093)

Used Car 0.012 -0.291*** -0.247*** -0.252*** -0.0106(0.416) (0.084) (0.081) (0.080) (0.044)

“Hard” Information

Obligor Credit Score 0.00400*** 0.000492** 0.000402* 0.000402** 0.000287(0.00113) (0.000222) (0.000211) (0.000194) (0.000197)

Payment-to-Income Ratio 0.139 -0.084 -0.108 -0.123 -0.024(1.097) (0.172) (0.131) (0.117) (0.148)

Original Interest Rate -13.85*** 0.122 0.113 -0.077 -1.490(3.109) (1.174) (1.162) (1.122) (1.073)

Original Loan Amount 3.22e-05* -1.63e-06 6.89e-07 1.26e-08 -3.31e-06*(1.63e-05) (1.80e-06) (1.20e-06) (1.03e-06) (1.73e-06)

Original Loan Term 0.0356** -0.00309 -0.00229 -0.00175 -0.00143(0.0148) (0.00192) (0.00161) (0.00147) (0.00185)

Geographic Location FE X X X X X

Origination Y-Q FE X X X X XSecuritization Y-Q FE X X X XVehicle Manufacturer FE XVehicle Make FE XLender Type FE X

Observations 4,535,240 4,535,240 4,533,466 4,535,240 4,535,240R2 0.938 0.996 0.996 0.996 0.997

Note: This table reports results from regressions of loan’s number of months for which it had been warehoused, onvarious independent variables of interest, controlling for public loan characteristics (“Original Interest Rate,” “OriginalLoan Amount,” “Original Loan Term,” “Obligor Credit Score,” and “Payment-to-Income Ratio”). “Income Verified”is a dummy variable that gives a value of 1 if the obligor has been verified of income at origination, and 0 otherwise.“Underwriting Exception” is a dummy variable that gives a value of 1 if the loan was an exception to underwritingcriteria and 0 otherwise. “Used Car” is a dummy variable that gives a value of 1 if the vehicle at origination is a usedcar and 0 otherwise. For fixed effects, lender types are grouped by captive auto finance companies, other non-bankfinance companies, and banks. All standard errors are clustered by ABS. Results using an alternative clustering byorigination year-quarter are similar (unreported). ***, **, and * denote statistical significance at the 1%, 5%, and10% level, respectively. Source: SEC form ABS-EE.

37

Table 3: Robustness Analysis Using New Car Loans

Dependent Variable: Warehoused Months(1) (2) (3) (4) (5)

“Soft” Information

Income Verified -0.405 0.0595 0.116 0.101 0.0363(0.441) (0.104) (0.0965) (0.0935) (0.113)

Underwriting Exception 1.873*** 0.338*** 0.197*** 0.176*** 0.330***(0.686) (0.0879) (0.0684) (0.0617) (0.0887)

“Hard” Information

Obligor Credit Score 0.00593*** 0.000410 0.000215 0.000240 0.000423(0.00144) (0.000298) (0.000239) (0.000223) (0.000287)

Payment-to-Income Ratio 0.0881 -0.0613 -0.0288 -0.0421 -0.0846(1.145) (0.200) (0.105) (0.0927) (0.171)

Original Interest Rate -16.00*** -1.602 -1.854 -2.073* -1.333(3.559) (1.202) (1.130) (1.099) (1.254)

Original Loan Amount 2.37e-05 -4.18e-06** -2.99e-07 -8.31e-07 -4.24e-06**(1.47e-05) (1.93e-06) (8.00e-07) (5.62e-07) (1.94e-06)

Original Loan Term 0.0720*** -0.000827 0.00134 0.00216 -0.000853(0.0213) (0.00232) (0.00159) (0.00137) (0.00233)

Geographic Location FE X X X X X

Origination Y-Q FE X X X X XSecuritization Y-Q FE X X X XVehicle Manufacturer FE XVehicle Make FE XLender Type FE X

Observations 2,286,735 2,286,735 2,286,479 2,286,735 2,286,735R2 0.925 0.996 0.996 0.996 0.996

Note: This table reports results from regressions of loan’s number of months for which it had been warehoused,on various independent variables of interest, controlling for public loan characteristics (“Original Interest Rate,”“Original Loan Amount,” “Original Loan Term,” “Obligor Credit Score,” and “Payment-to-Income Ratio”), basedon a sub-sample of loans originated on new cars. “Income Verified” is a dummy variable that gives a value of 1 if theobligor has been verified of income at origination, and 0 otherwise. “Underwriting Exception” is a dummy variablethat gives a value of 1 if the loan was an exception to underwriting criteria and 0 otherwise. “Used Car” is a dummyvariable that gives a value of 1 if the vehicle at origination is a used car and 0 otherwise. For fixed effects, lender typesare grouped by captive auto finance companies, other non-bank finance companies, and banks. All standard errorsare clustered by ABS. Results using an alternative clustering by origination year-quarter are similar (unreported).***, **, and * denote statistical significance at the 1%, 5%, and 10% level, respectively. Source: SEC form ABS-EE.

38

Table 4: Robustness Analysis Using Post-2017 Originated Loans

Dependent Variable: Warehoused Months(1) (2) (3) (4) (5)

“Soft” Information

Income Verified 0.571* 0.277*** 0.292*** 0.299*** 0.0864(0.341) (0.103) (0.0990) (0.0952) (0.0876)

Underwriting Exception 2.478*** 0.644*** 0.587*** 0.582*** 0.462***(0.397) (0.142) (0.131) (0.130) (0.134)

Used Car -0.521* -0.261*** -0.248*** -0.254*** 0.0423(0.277) (0.0911) (0.0867) (0.0855) (0.0522)

“Hard” Information

Obligor Credit Score 0.00487*** 0.000513** 0.000399* 0.000380* 0.000432*(0.000962) (0.000235) (0.000223) (0.000225) (0.000231)

Payment-to-Income Ratio -0.607 -0.192 -0.237 -0.228 -0.0907(0.627) (0.182) (0.146) (0.144) (0.169)

Original Interest Rate -16.24*** -1.332 -1.352 -1.434 -2.619***(2.186) (1.166) (1.126) (1.089) (0.952)

Original Loan Amount 1.52e-05* 1.15e-07 1.39e-06 9.39e-07 -2.40e-06(8.25e-06) (2.01e-06) (1.38e-06) (1.40e-06) (1.86e-06)

Original Loan Term 0.00725 -0.00197 -0.00166 -0.00134 -0.000104(0.00670) (0.00176) (0.00131) (0.00125) (0.00179)

Geographic Location FE X X X X X

Origination Y-Q FE X X X X XSecuritization Y-Q FE X X X XVehicle Manufacturer FE XVehicle Make FE XLender Type FE X

Observations 2,658,123 2,658,123 2,657,791 2,658,123 2,658,123R2 0.439 0.924 0.925 0.925 0.930

Note: This table reports results from regressions of loan’s number of months for which it had been warehoused,on various independent variables of interest, controlling for public loan characteristics (“Original Interest Rate,”“Original Loan Amount,” “Original Loan Term,” “Obligor Credit Score,” and “Payment-to-Income Ratio”), basedon a sub-sample of loans originated after 2017. “Income Verified” is a dummy variable that gives a value of 1 if theobligor has been verified of income at origination, and 0 otherwise. “Underwriting Exception” is a dummy variablethat gives a value of 1 if the loan was an exception to underwriting criteria and 0 otherwise. “Used Car” is a dummyvariable that gives a value of 1 if the vehicle at origination is a used car and 0 otherwise. For fixed effects, lender typesare grouped by captive auto finance companies, other non-bank finance companies, and banks. All standard errorsare clustered by ABS. Results using an alternative clustering by origination year-quarter are similar (unreported).***, **, and * denote statistical significance at the 1%, 5%, and 10% level, respectively. Source: SEC form ABS-EE.

39

Tab

le5:

Au

toR

ecal

lsan

dA

ver

age

Imp

acts

,20

09-2

018

Au

tom

aker

Tot

alN

um

ber

Aver

age

Aff

ecte

dA

vera

geS

cale

dA

ver

age

Aver

age

Ave

rage

ofR

ecal

lsU

nit

sA

ffec

ted

Un

its

New

sV

olu

me

Sen

tim

ent

Rec

all

Imp

act

Tes

la9

21,4

960.

1583

7.73

49.0

12.9

0T

oyot

a10

026

5,08

10.

0110

32.8

948

.39

0.2

9G

M14

416

4,95

30.

0017

41.3

351

.15

0.2

2H

ond

a87

281,

405

0.02

392.

8650

.01

0.1

8V

olksw

agen

121

39,9

440.

0198

0.46

49.3

50.1

7N

issa

n80

146,

558

0.01

490.

9153

.39

0.1

3F

CA

170

131,

784

0.01

734.

7448

.25

0.1

2F

ord

140

162,

638

0.01

955.

6552

.22

0.1

0H

yu

nd

ai67

120,

381

0.02

213.

552

.62

0.0

9K

ia29

102,

054

0.03

95.4

551

.58

0.0

9S

uzu

ki

1434

,524

0.06

60.9

452

.63

0.0

8M

itsu

bis

hi

4840

,867

0.03

90.5

751

.17

0.0

7A

ston

Mar

tin

71,

411

0.05

53.9

548

.62

0.0

6B

MW

100

47,6

840.

0126

0.25

53.9

90.0

6M

clar

en2

241

0.11

31.9

861

.42

0.0

6T

ata

547,

470

0.01

343.

2455

.40.0

6M

azd

a42

104,

670

0.03

61.5

148

.83

0.0

4D

aim

ler

289

7,81

60.

0041

6.39

54.9

80.0

1V

olvo

6811

,257

0.01

48.4

456

.95

0.0

1S

ub

aru

3010

2,76

00.

048.

0745

.82

0.0

0

Tot

al1,

601

Note

:W

eco

llec

tdata

on

all

auto

reca

lls

from

2009

to2018

for

all

auto

maker

sw

hose

cars

app

ear

on

our

main

data

set.

”A

ffec

ted

Unit

s,”

”Sca

led

Aff

ecte

dU

nit

s,”

”N

ews

Volu

me,

””Sen

tim

ent,

”and

”R

ecall

Impact

”are

inav

erages

.A

ver

age

Aff

ecte

dU

nit

sare

the

aver

age

num

ber

of

cars

aff

ecte

dby

the

auto

maker

’sva

rious

reca

llin

stance

s.A

ver

age

Sca

led

Aff

ecte

dU

nit

sadju

stth

eim

pact

of

each

reca

llby

scaling

the

aff

ecte

dunit

sof

reca

lled

veh

icle

sby

the

esti

mate

dnum

ber

of

the

veh

icle

sof

reca

lled

model

-yea

rth

at

rem

ain

on

the

road

at

the

tim

eof

the

reca

ll.

We

win

sori

zeobse

rvati

ons

at

bott

om

and

top

1%

of

scale

daff

ecte

dunit

sto

excl

ude

outl

iers

.A

ver

age

New

sV

olu

me

and

Aver

age

Sen

tim

ent

Sco

reea

chco

me

from

Rav

enpack

New

sA

naly

tics

’“A

ggre

gate

Even

tV

olu

me”

and

“A

ggre

gate

Even

tSen

tim

ent,

”co

ndit

ional

on

the

rele

vance

of

the

new

sev

ent

bei

ng

larg

erth

an

75

(on

a1-1

00

scale

);ea

chm

easu

reth

evolu

me

of

even

tsre

port

edin

financi

al

new

sm

ediu

mand

the

share

of

posi

tive

even

ts,

on

a91-d

ayro

llin

gbasi

s.A

ver

age

Rec

all

Impact

isth

eav

erage

of

the

logari

thm

of

Rec

all

Impact

defi

ned

asRecallImpact

it=ln

(ScaledAffectedUnits i

t∗NewsVolume i

t/Sentiment i

t).

Sourc

e:N

ati

onal

Hig

hw

ayT

raffi

cSafe

tyA

dm

inis

trati

ons

Offi

ceof

Def

ects

Inves

tigati

on,

IHS

Mark

it’s

New

Veh

icle

Reg

istr

ati

on,

Polk

’sV

ehic

les

inO

per

ati

on,

and

Rav

enpack

New

sA

naly

tics

.

40

Table 6: Exogenous Variations in Collateral/Borrower Quality: Auto Recalls and “Super Sales”

Dependent Variable: Warehoused Months

Panel A: Originated Prior to Recall Panel B: Super Sales(1) Non-Automakers (2) Automakers (3) All issuers

1-Mo Post-Orig Recall Impact -0.0214** -0.00698(0.00998) (0.0112)

End-of-Year Origination -1.705***(0.0389)

Holiday-Sale Origination -0.390***(0.0628)

Income Verified 0.171*** 0.0812 0.0227(0.0562) (0.0635) (0.0422)

Underwriting Exception 0.895*** 0.118*** 0.0294**(0.323) (0.0433) (0.0144)

Used Car -0.176* 0.0249 0.0371***(0.0870) (0.0230) (0.0128)

Obligor Credit Score 0.000145 -4.84e-05 0.000119*(0.000236) (0.000195) (6.87e-05)

Payment-to-Income Ratio -0.239* 0.0569 -0.0306(0.129) (0.0823) (0.0354)

Original Interest Rate 0.971 -1.041** -0.972***(1.052) (0.472) (0.234)

Original Loan Amount 5.17e-08 -7.96e-07 4.68e-07**(1.60e-06) (5.46e-07) (2.22e-07)

Original Loan Term -0.00122 0.000514 0.00123***(0.00255) (0.00127) (0.000464)

Geographic Location FE X X XOrigination Y-Q FE X X XSecuritization Y-Q FE X X XVehicle Manufacturer FE X X X

Observations 2,213,065 2,320,401 4,533,466R2 0.997 0.996 0.998

Note: This table reports results from robustness analyses using auto recalls and super sales. In Panel A, “1-MoPost-Origination Recall Impact” is the logarithm of Recall Impact of the loan vehicle’s automaker k during onemonth after the loan’s origination, where Recall Impact is defined as Recall Impactkt = ln(ScaledAffectedUnitskt ∗NewsV olumekt/Sentimentkt) (see Data Appendix for more details). Column (1) shows results from a sub-sampleof lenders that are not auto manufacturers and Column (2) shows results from the automakers. In Panel B, “End-of-Year Origination” and “Holiday-Sale Origination” each indicate whether the loan were originated at the end of ayear (defined as in December) and or in months of May (Memorial Day sale) or November (Black Friday). “IncomeVerified” is a dummy variable that gives a value of 1 if the obligor has been verified of income at origination, and0 otherwise. “Underwriting Exception” is a dummy variable that gives a value of 1 if the loan was an exceptionto underwriting criteria and 0 otherwise. “Used Car” is a dummy variable that gives a value of 1 if the vehicle atorigination is a used car and 0 otherwise. All standard errors are clustered by ABS. Results using an alternativeclustering by origination year-quarter are similar (unreported). ***, **, and * denote statistical significance at the1%, 5%, and 10% level, respectively. Source: SEC form ABS-EE, National Highway Traffic Safety AdministrationsOffice of Defects Investigation, IHS Markit’s New Vehicle Registration, Polk’s Vehicles in Operation, and RavenpackNews Analytics.

41

Table 7: Exogenous Variations in Borrower Quality: Model Updates

Dependent Variable: Warehoused Months(1) Originated -1 Mo (2) Originated -3 Mo (3) Originated -6 Mo

Model Update -0.140*** -0.0306 -0.0268(0.0454) (0.0393) (0.0340)

Income Verified 0.213** 0.213** 0.213**(0.0901) (0.0901) (0.0901)

Underwriting Exception 0.376*** 0.376*** 0.376***(0.0901) (0.0901) (0.0901)

Used Car -0.209*** -0.210*** -0.210***(0.0755) (0.0755) (0.0755)

Obligor Credit Score 0.000289 0.000289 0.000289(0.000206) (0.000206) (0.000206)

Payment-to-Income Ratio -0.0995 -0.0996 -0.0995(0.127) (0.127) (0.127)

Original Interest Rate -0.0750 -0.0709 -0.0718(1.160) (1.160) (1.160)

Original Loan Amount 5.17e-07 5.44e-07 5.42e-07(1.07e-06) (1.06e-06) (1.06e-06)

Original Loan Term -0.00228 -0.00228 -0.00229(0.00161) (0.00160) (0.00161)

Geographic Location FE X X XOrigination Y-Q FE X X XSecuritization Y-Q FE X X XVehicle Manufacturer FE X X X

Observations 2,748,388 2,748,388 2,748,388R2 0.996 0.996 0.996

Note: This table reports results from robustness analyses using model updates. ”Model Update” is a dummy variablethat gives a value of 1 if the loan was originated 1, 3, and 6 months (Columns (1)-(3), respectively) prior to a modelupdate on the model of the car associated with the loan, and 0 otherwise. “Income Verified” is a dummy variable thatgives a value of 1 if the obligor has been verified of income at origination, and 0 otherwise. “Underwriting Exception”is a dummy variable that gives a value of 1 if the loan was an exception to underwriting criteria and 0 otherwise.“Used Car” is a dummy variable that gives a value of 1 if the vehicle at origination is a used car and 0 otherwise. Allstandard errors are clustered by ABS. Results using an alternative clustering by origination year-quarter are similar(unreported). ***, **, and * denote statistical significance at the 1%, 5%, and 10% level, respectively. Source: SECform ABS-EE and Kelley Blue Book.

42

Tab

le8:

An

alysi

sof

Pos

t-S

ecu

riti

zati

onL

oan

Per

form

ance

Pan

elA

:Pr(

Ch

arge

-off

)P

anel

B:

Char

ged

-off

Pri

nci

pal

Am

ou

nt

($)

Pos

t-S

ecu

riti

zati

on(1

)+

3-M

onth

(2)

+6-

Mon

th(3

)+

12-M

onth

(4)

+3-

Mon

th(5

)+

6-M

onth

(6)

+12-M

onth

Sikto

-3.8

04**

*-4

.819

***

-6.6

93**

*-1

1154

.640

***

-110

44.3

20***

-11073.8

90***

(0.0

09)

(0.0

11)

(0.0

17)

(34.

796)

(31.8

89)

(33.1

08)

Ob

ligo

rC

red

itS

core

0.00

1709

***

0.00

2714

***

0.00

3956

***

3.18

3915

***

3.83

4559***

4.2

91813***

(0.0

0002

8)(0

.000

033)

(0.0

0004

7)(0

.118

503)

(0.1

17162)

(0.1

27389)

Pay

men

t-to

-In

com

eR

atio

2.12

9***

1.98

1***

2.14

1***

8521

.455

***

7425

.434***

6680.9

06***

(0.0

40)

(0.0

47)

(0.0

66)

(180

.635

)(1

75.3

64)

(189.4

89)

Ori

gin

alIn

tere

stR

ate

-14.

655*

**-1

2.65

9***

-11.

483*

**-4

0300

.000

***

-303

76.0

70***

-24774.2

20***

(0.0

59)

(0.0

68)

(0.0

94)

(224

.950

)(2

23.9

57)

(242.7

52)

Ori

gin

alL

oan

Am

ount

-1.9

5E-0

5***

-1.8

1E-0

5***

-1.8

5E-0

5***

-6.5

8E-0

2***

-3.6

9E-0

2***

-3.0

1E

-02***

(1.8

4E-0

7)(2

.18E

-07)

(3.1

4E-0

7)(7

.56E

-04)

(8.0

1E

-04)

(8.6

8E

-04)

Ori

gin

alL

oan

Ter

m0.

0233

7***

0.03

308*

**0.

0403

7***

58.3

9854

***

82.0

2102***

75.8

0086***

(0.0

0021

)(0

.000

25)

(0.0

0036

)(0

.948

08)

(0.9

2599)

(0.9

8692)

Geo

grap

hic

Loca

tion

FE

XX

XX

XX

Ori

gin

atio

nY

-QF

EX

XX

XX

XS

ecu

riti

zati

onY

-QF

EX

XX

XX

XV

ehic

leM

anu

fact

ure

rF

EX

XX

XX

X

Ob

serv

atio

ns

2,58

8,32

82,

047,

686

1,21

1,69

42,

475,

795

2,01

2,4

56

1,2

09,8

21

(Pse

ud

o)R

20.

270.

310.

390.

230.

21

0.2

6

Note

:T

his

table

rep

ort

sre

sult

sfr

om

logit

regre

ssio

ns

of

post

-sec

uri

tiza

tion

pro

babilit

yof

def

ault

,re

poss

essi

on,

and

charg

e-off

onSik

to,

the

loan-l

evel

sum

mary

stati

stic

on

signaling.

We

rep

ort

coeffi

cien

tsusi

gn

3-,

6-,

and

12-m

onth

post

-sec

uri

tiza

tion

loan

outc

om

es.

***,

**,

and

*den

ote

stati

stic

al

signifi

cance

at

the

1%

,5%

,and

10%

level

,re

spec

tivel

y.Sourc

e:SE

Cfo

rmA

BS-E

E.

43

Table 9: Distribution of Warehoused Months by Lender/Issuer Type

Mean Stdev N

All issuers 15 16 5,930,935Captive auto 16 15 3,470,732Banks 13 18 1,798,168

Banks without Santander 19 16 621,601Santander 10 19 1,176,567

Non-banks 12 19 662,035

Note: This table shows the average and the standard deviation of warehoused months by each lender/issuer type— captive auto finance companies (BMW, Daimler, Ford, GM, Honda, Hyundai, Nissan, Toyota, and Volkswagen),other non-bank finance companies (Carmax), banks exluding Santander (Ally, California Republic, Fifth Third, andUSAA), and Santander. Source: SEC form ABS-EE.

44

Table 10: Asymmetric Information and Signaling by Lender Type

Dependent Variable: Warehoused Months

(1) Banks (2) Captive Autos (3) Other Non-banks“Soft” Information

Income verified 0.125*** 0.0816 -0.0147(0.0416) (0.0634) (0.0458)

Underwriting exception 0.571* 0.118***(0.281) (0.0433)

Used 0.0665* 0.0244 0.0358(0.0367) (0.0231) (0.0202)

“Hard” Information

Obligor Credit Score -0.000212 -5.00e-05 -0.000230***(0.000306) (0.000193) (3.92e-05)

Payment-to-Income Ratio 0.0601 0.0591 -0.0735**(0.138) (0.0828) (0.0291)

Original Interest Rate -1.359 -1.055** 0.133(1.118) (0.477) (0.117)

Original Loan Amount 1.15e-06 -8.22e-07 -4.48e-08(1.72e-06) (5.56e-07) (2.24e-07)

Original Loan Term -0.00251 0.000598 0.00163***(0.00233) (0.00124) (0.000308)

Geographic Location FE X X XOrigination Y-Q FE X X XSecuritization Y-Q FE X X XVehicle Manufacturer FE X X X

Observations 1,613,667 2,320,401 599,398R2 0.997 0.996 0.999

Note: This tables shows the results of baseline estimation by each lender/issuer type — captive auto finance companies(BMW, Daimler, Ford, GM, Honda, Hyundai, Nissan, Toyota, and Volkswagen), other non-bank finance companies(Carmax), and banks (Ally, California Republic, Fifth Third, Santander, and USAA). “Income Verified” is a dummyvariable that gives a value of 1 if the obligor has been verified of income at origination, and 0 otherwise. “UnderwritingException” is a dummy variable that gives a value of 1 if the loan was an exception to underwriting criteria and0 otherwise. “Used Car” is a dummy variable that gives a value of 1 if the vehicle at origination is a used carand 0 otherwise. All standard errors are clustered by ABS. Results using an alternative clustering by originationyear-quarter are similar (unreported). ***, **, and * denote statistical significance at the 1%, 5%, and 10% level,respectively. Source: SEC form ABS-EE.

45

Table 11: Summary Statistics of ABS-level Characteristics

Mean Stdev Min Max

Average ABS Spread (%) -0.06 0.38 -0.83 0.85Face Amount ($ Mil) 1,200 310 421 2,230Number of Loans 66,804 19,955 20,737 109,465Number of Tranches 7 1 4 8

Average Skt (Weighted) -0.01 0.14 -0.25 0.29Average Warehoused Months (Weighted) 14.16 5.92 2.16 23.91Average Maturity (Weighted) 57 9 28 70Average Credit Score (Weighted) 700 70 567 765

Total Number of ABS 87

Note: We calculate the average spread by taking the unweighted average of coupon rates of different tranches withineach pool and subtracting the 5-year Treasury yield. Skt is a measure of weighted average signaling of loans in poolk, defined as follows:

Skt =∑i∈k

ωiktSikto

=∑i∈k

ωikt × (β11 × IncomeV erifiedikto + β12 × Exceptionikto + β13 × UsedCarikto)

where ωikt is the origination loan amount share of loan i in pool k. We use the estimated coefficients, β11, β12, andβ13, from the baseline estimation in Table 2 Column (3). Higher Skt implies that the pool contains more unobservablygood quality loans — in the form of loans that are verified of income, that have exceptionally good credit factors, orthat are made on new cars. Source: Bloomberg ABS Backoffice, SEC form ABS-EE, and St. Louis FRED database.

46

Table 12: Pricing Analysis of Auto ABS

Dependent Variable: Average Spread(1) (2) (3) (4)

Avg(Sikto) (Equally, Skt) -1.065**(0.494)

Med(Sikto) -0.200(0.274)

Min(Sikto) -0.316**(0.144)

Stdev(Sikto) -1.722**(0.803)

Face Amount ($; Log) -0.209 -0.132 -0.150 -0.197(0.249) (0.274) (0.246) (0.244)

Number of Loans (Log) -0.0612 -0.254 -0.136 -0.0744(0.234) (0.224) (0.219) (0.236)

Average Credit Score (Weighted) -0.00116* -0.000840 -0.00123* -0.00154**(0.000679) (0.000696) (0.000688) (0.000780)

Number of Tranches -0.0166 0.0195 0.0367 0.00112(0.0471) (0.0438) (0.0406) (0.0452)

Securitization Y-Q FE X X X XLender Type FE X X X XObservations 87 87 87 87R2 0.711 0.685 0.701 0.705

Note: We calculate the average spread by taking the (unweighted) average of coupon rates of different tranches withineach pool and subtracting the 5-year Treasury yield. Sikto is a measure of signaling on loan i in pool k, defined asfollows:

Sikto = β11 × IncomeV erifiedikto + β12 × Exceptionikto + β13 × UsedCarikto

We use the estimated coefficients, β11, β12, and β13, from the baseline estimation in Table 2 Column (3). HigherSikto implies that the loan i is of unobservably good quality than other loans by being verified of income or havingexceptionally good credit factors, or by being made on a new car. Columns (1)-(4) each takes the weighted average,median, minimum, and standard deviation of Sikto within the pool. For fixed effects, lender types are grouped bycaptive auto finance companies, other non-bank finance companies, banks excluding (subprime-focused) Santander,and Santander. All standard errors are bootstrapped. ***, **, and * denote statistical significance at the 1%, 5%,and 10% level, respectively. Source: Bloomberg ABS Backoffice, SEC form ABS-EE, and St. Louis FRED database.

47

Appendix

A. Representations and Warranties: “No Adverse Selection” in US Auto ABS

Market

In August 2014, the SEC amended Rule 17g-7 so as to require nationally recognized statistical rating

organizations (NRSROs), when assigning a credit rating to an asset-backed security, to disclose:

“(1) the representations, warranties, and enforcement mechanisms available to investors which were

disclosed in the prospectus, private placement memorandum, or other offering documents for the

asset-backed security and that relate to the asset pool underlying the asset-backed security; and

(2) how they differ from the representations, warranties, and enforcement mechanisms in issuances

of similar securities.”

One of the representations and warranties typically addressed in US auto ABS deals is that

there was no adverse selection in the creation of the pool of loans. Fitch Ratings describes it as:

“No Adverse Selection: No selection procedures believed to be adverse to the noteholders have

been utilized in selecting the receivables from other receivable of the sponsor that meet the criteria

specified in the transaction documents.”14

However, the SEC’s rule only requires the rating agencies to disclose information on reps and

warranties based on the offering and transaction documents provided by the ABS issuer to the

rating agency. An example of this disclosure is:

..Moody’s has not undertaken any other investigation into the accuracy of the issuer’s statement.

In rating the Transaction, Moody’s evaluates the representations, warranties and enforcement

mechanisms contained in the offering and transaction documents solely as and to the extent

described in its rating criteria. Further, Moody’s rating may depend significantly on factors

other than such representations, warranties and enforcement mechanisms. Moody’s does not in

this 17g-7 Report provide any opinion or recommendation as to the adequacy or effectiveness

of the representations, warranties and enforcement mechanisms described herein (whether with

respect to the Transaction or the Benchmark). Investors must conduct their own analysis of

the adequacy and effectiveness, and of the legal and other implications, of the representations,

warranties and enforcement mechanisms in the Transaction.15

Because rating agencies do not necessarily conduct their own analysis to confirm on reps and

warranties, it ultimately falls upon the due diligence of deal issuers to ensure appropriateness of

reps and warranties. A part of such reps and warranties-related information disclosed by the issuer

to the rating agency can come from the deal prospectus. In addition to summary statistics of the

pool as in Figure 4, the following excerpt from the prospectus of Honda Auto Receivables 2018-

1 Owner Trust shows an example discussion of pool characteristics that can feed into the above

14FitchRatings Special Report, 2016, “Representations, Warranties and Enforcement Mechanisms in Global Struc-tured Finance Transactions”.

15Moody’s SEC Rule 17g-7 Report of R&Ws: Ford Credit Auto Owner Trust 2019-A Deal v1.1 Compared To AutoLoans v3.0 (highlights added by authors).

48

disclosure by rating agencies:

Characteristics of the Receivables The Receivables to be held by the trust will be

selected from those motor vehicle retail installment sale contracts in AHFC’s portfo-

lio that meet several criteria as of the Cutoff Date. These criteria provide that each

Receivable:

• was originated by a dealer located in the United States and the Obligor

is not (according to the records of AHFC) a federal, state or local gov-

ernmental entity;

• has a contractual annual percentage rate specified in the promissory note

associated with each Receivable (which we refer to in this prospectus as

the “APR”) of at least 0.50%;

• has an original term to maturity of not more than 72 months;

• is not more than 30 days past due;

• has been entered into by an Obligor that was not in bankruptcy proceed-

ings or is bankrupt or insolvent (according to the records of AHFC);

• is attributable to the purchase of a new or used Honda or Acura auto-

mobile or light-duty truck and is secured by that automobile

• provides for the related monthly payment on a Financed Vehicle owed

by the related Obligor (each such payment, a “Scheduled Payment”)

according to the simple interest method (as described below); and

• except as otherwise permitted under the sale and servicing agreement,

provides for scheduled monthly payments that fully amortize the amount

financed by such Receivable over its original term (except that the first

or last payment in the life of the Receivable may be minimally different

from the level payment).

Payments on Receivables using the “simple interest method” will be applied first to

interest accrued through the date immediately preceding the date of payment and then

to unpaid principal. Accordingly, if an Obligor pays an installment before its due date,

the portion of the payment allocable to interest for the payment period will be less than

if the payment had been made on the due date, the portion of the payment applied to

reduce the Principal Balance will be correspondingly greater, and the Principal Balance

will be amortized more rapidly than scheduled.

The ability of the servicer to make modifications on the Receivables is not expected to

have a material impact on the distributions on the notes described in this prospectus.

For a description of the servicer’s ability to make modifications to the Receivables, see

“Description of the Transfer and Servicing Agreements—Servicing Procedures.”

49

No selection procedures believed to be adverse to the noteholders will be utilized in

selecting the Receivables from qualifying retail installment sale contracts or from the

receivables in the pool. For a description of AHFC’s loss and delinquency experience

on its managed pool portfolio, see “The Sponsor, Originator, Servicer and Administra-

tor—Servicing Experience.”

B. Construction of “Recall Impact”

Data on vehicle recalls is from the National Highway Transportation Safety Administration’s Office

of Defects Investigation.16

To measure the impact of each recall, we scale them using two datasets. The first is Polk’s

Vehicles in Operation data which give the stock of vehicles in operation by model-year for years

from 2000 onwards. The second dataset is IHS Markit’s New Vehicle Registration which has

information on flows at the make/model/model-year level.

We combine this information to create a variable named Scaled Affected Units. It is constructed

as follows. The stock of vehicles-in-operation at the model-year level is available across all makes

and models. Using this data, we calculate the share of all vehicles of a given model-year that are

still on the road in subsequent years; this proxies for a model-year “survival rate”. Next, we apply

these survival rates to more granular new vehicle registration data at the manufacturer level to

estimate the stock of vehicles-in-operation for a given manufacturer in each year of the sample.

Finally, we scale the raw units recalled by the estimated stock of vehicles-in-operation.

Putting these steps together, we define and measure the “Recall Impact” by multiplying the

Scaled Affected Units by the News Volume and dividing by Sentiment. We obtain data on News

Volume and Sentiment from Ravenpack News Analytics’ “Aggregate Event Volume” and “Aggre-

gate Event Sentiment,” conditional on the “Relevance” of the news event being larger than 75 (on

a 1-100 scale). Aggregate Event Volume and Aggregate Event Sentiment each measure the volume

of events reported in financial news medium and the share of positive sentiment events, on a 91-

day rolling basis. Ravenpack determines scores on positive sentiment by “systematically matching

stories typically categorized by financial experts as having short-term positive or negative financial

or economic impact.” (Source: Ravenpack News Analytics User Guide)

16A detailed description and links to download the data are available at https://catalog.data.gov/dataset/nhtsas-office-of-defects-investigation-odi-recalls-recalls-flat-file.

50